Why You Cannot See the Install Yet
Your data lives in different rooms. Project leads and plans sit in market databases. Accounts and opportunities live in CRM. Quotes and shipments hide in distributor or dealer systems. Website forms, sample requests, and document downloads sit in web analytics. None of it was designed to talk to each other, so the trail from specification to submittal to install is broken.
The market itself moves before spend shows up. Public data shows U.S. construction spending near two trillion dollars, but what matters to your sellers are the early tells like planning activity and specs that harden into buys. The Dodge Momentum Index is a leading signal for nonresidential spend by roughly a year to eighteen months, which is why tying your internal signals to it pays off in time to act. See the latest DMI note for March 2026 growth and its lead-time explanation here. Total put-in-place estimates for December 2025 sat near $2.17 trillion seasonally adjusted, which frames the opportunity size per the Census release. That scale is motivating, and also messy to follow in real life.
Treat It Like a Demand Signal Spine
Think of a simple spine that all signals can attach to. The vertebrae are project, firm, location, CSI section, and time. Deterministic matches come from shared IDs, emails on RFIs, or PO numbers. Probabilistic matches use fuzzy company names, addresses, spec section overlap, and contact graphs across GCs, subs, and architects. The goal is not a perfect master record. It is a working hypothesis that a specific project will probably install your SKU family within a time window.
Entity resolution sounds exotic but is just careful record linkage with confidence scoring. Start with the minimum viable spine. One project source, one CRM export, one distributor shipment file, one web analytics export. Weekly batch is fine. Your first win is a ranked list of projects with a reason to call, not a lake full of gold.
What Counts As a High-Quality Signal
A specification mention with product attributes that match your catalog is strong, especially if the section number and performance language narrow the field. A reviewed submittal is stronger still because it confirms intent and timing. Many public owners and large programs follow standard submittal procedures like those documented in Division 01, which creates consistent breadcrumbs for AI to read. See examples of Submittal Procedures in the Unified Facilities Guide Specifications index here.
Web signals help when they tie back to the same spine. Sample requests, datasheet downloads, and spec-builder tool usage often correlate with active design phases. If you need to stitch offline interactions, Google’s Measurement Protocol for GA4 allows sending offline events so you can align site behavior with CRM activity per Google’s developer guide. Use that carefully and avoid personally identifiable information.
A Practical Model For “Probable Install”
Your model does not need deep magic. Gradient-boosted trees or a logistic regression with well-chosen features will outperform a hand-scored lead sheet. Useful features include spec strength and novelty of language, distance from design release to bid date, the network of firms tied to similar past installs, shipment proximity and quantity patterns, and recent on-site activity in project updates. Text embeddings can turn spec and submittal wording into numeric signals without rewriting your systems.
The target is a simple probability that a project will result in your product being installed within a given horizon. Express results as a ranked queue for sales and technical services. Add short explanations so humans can sanity-check the call plan. Trust grows when the why is clear.
Data You Actually Need To Start
Aim for four inputs before you boil the ocean. Project plans and specs with section numbers. CRM opportunities with account, project name, and close dates. Distributor or dealer shipments with city, project or job name if available, and SKU family. Web analytics events for forms, samples, and document downloads keyed by company domain or campaign ID. That is enough to build the first spine and score.
Keep the join keys boring. Standardize company and project names. Normalize addresses. Map product families to CSI sections used by the teams that buy from you. Document the matching rules so commercial and IT can audit them.
Guardrails and Data Responsibilities
AI that touches customer and project data needs risk checks. The NIST AI Risk Management Framework and its Generative AI Profile offer practical controls for mapping risks to mitigations and for documenting system intent, data provenance, and evaluation methods. Use these as a common language across sales, IT, and legal see NIST’s profile.
Privacy and procurement rules vary by owner and project. Keep web analytics free of disallowed personal data. Maintain an opt-out path for marketing contacts. Log which datasets feed each score and who viewed the output. If requirements change, document the change date inside the scoring pipeline so audits are possible later.
Sales Workflow That Puts AI To Work
Publish a weekly ranked list of projects by territory with three fields sellers care about. What happened. Why it matters. Who to call next. Example entries might be “Spec tightened from ‘or equal’ to named attributes,” “Submittal returned approved as noted,” or “Distributor shipment to same GC within 10 miles last week.”
Technical services can pair this with prepared responses and comparison sheets that speak to the exact section language. Marketing can suppress broad campaigns to accounts already deep in a live spec to avoid noise. Leadership can track coverage of the top one hundred projects rather than raw lead counts.
Common Failure Modes To Avoid
Chasing every spec mention evenly wastes time. Weight submittals and firm networks higher than casual downloads. Over-indexing on shipments without a project join leads to false positives on maintenance or replacement jobs. Ignoring owner type and funding can skew timing. Treat alternates and value engineering as explicit features, not afterthoughts.
The last trap is perfectionism. You will not get every join right. A spine with 80 percent precision that routes the next ten calls better than last week is already compounding value.
How To Get Moving In 2026
Pick one product family and one region. Connect two data sources first, not six. Stand up a weekly batch. Score every project and route the top twenty to sales with a one-line reason. Expand data coverage only when sellers report that the top of the list feels obviously better than before. That is your north star.
If you want a public benchmark to track against your internal signals, tie your project list to a leading indicator. The DMI is updated monthly and explicitly leads nonresidential spend by about a year, which gives you an external sanity check on your pipeline mix see March 2026 DMI context. For scale context on where spend actually lands, keep an eye on the Census’ put-in-place series latest release linked above.
